ChatGPT Almost Became a Meeting Bot. Its Co-Creator Is Now Rearranging Atoms

Elad Gil introduces Liam Fedus as a co-creator of ChatGPT, former VP of post-training at OpenAI, and a Google Brain veteran who worked on one of the first trillion-parameter models. Fedus starts further back. He was a physics major who spent time on dark matter research, including an apparatus that was directionally sensitive to incoming dark matter. Sarah Guo cuts in with a question that has become a running joke in the field: why are so many AI leaders physicists? She names Dario Amodei at Anthropic and Adam Brown at Google.

Fedus says physics is a principled, hard-nosed way to think, and computer science currently offers unusual leverage. After the discovery of the Higgs, high-energy physics became bottlenecked on the next apparatus for the energy frontier. Physicists looked at their skill set, looked at progress elsewhere, and decided they could contribute more in AI. Guo notes that string theorists and people working on black holes have been moving in. It almost feels, she says, like recreating a Manhattan Project, except the search is for different forms of intelligence.

In graduate school he kept gravitating toward machine-learning problems such as particle reconstruction, then concluded that pushing the ML frontier meant going into computer science. He landed at Google Brain in 2016 and 2017, overlapping with the first year of residents, in the stretch that produced distributed training strategies, mixture of experts, and the transformer. He calls it a Cambrian era: small collaborations, a handful of GPUs, and a lot of diversity and entropy in the research. For years afterward he stayed on architecture, pushing sparsity so large models could be served more cheaply.

By late 2022 the technology was finally compelling enough to turn into products, and Fedus left for OpenAI with other former Googlers. The brief was to productionize GPT-4, which was already pretrained and had some rough post-training. The group spitballed a writing bot, a coding bot, and, among the least interesting ideas, a meeting bot that would sit in Google Meet, take notes, and email to-dos after the call.

John Schulman was opinionated: keep it general, do a chatbot. That became a large part of the effort for those few months, and ChatGPT was the result. Gil, who had been investing in the area before the launch, says the technology had felt like a secret until the product shipped and everyone suddenly realized it was available. He calls that release the starting gun of public awareness of the current AI wave.

Gil asks why, with language transforming so much of software, Fedus went back to materials and atoms. Fedus says connecting these systems to the physical world was inevitable. The opinion he and the other Periodic founders held is that science will not see the same acceleration unless models interface with reality. "Science ultimately isn't sitting in a room thinking really hard," he says. You have to run experiments, learn from them, and touch the world.

ChatGPT in late 2022 was important, but still far too weak. Periodic could not have been built on that era's technology. Over the next few years, models improved, reasoning and test-time inference became important, error correction and tool use got more reliable, and coding agents arrived. Those, Fedus argues, were foundational technologies required to connect AI to the physical world. It was simply not possible with the systems of 2022.

Gil points out the other missing piece: data. Language models trained on the internet; atoms have no equivalent corpus. Fedus answers with two sources, physics simulation and experiment, plus a warning. Machine-learning systems are good on the data and tasks they were trained on. There is a mythology around AGI, ASI, and recursive self-improvement, but even increasingly powerful systems become limited without access to the raw data needed to make informed decisions.

Periodic does not reinvent English or code. It uses a mix of open and closed models, spends zero effort improving coding models, and treats Codex and Claude Code as accelerators, concentrating ML work where existing frontiers are not good enough. The company leans on the order of tens of trillions of tokens that went into open-source models for a foundational picture of the world. Once it moves into specific chemical spaces, sample efficiency is high because the system is not a randomly initialized net. It arrives with a strong prior from papers and the internet.

That prior is still insufficient. One engineer looked at a reported material property extracted from the literature and found the values spanned many orders of magnitude. Train on that, Fedus says, and the best you can do is model the distribution; you are no closer to ground truth. Experimental data supplies the grounding. What matters is not a pool of numbers but an interactive closed loop: look for aberrations and patterns, check consistency with simulation and literature, then drive the next experiments.

Guo, who used to work as a biologist, raises AlphaFold. Crystal structures used to take years, if they happened at all, because a protein might never crystallize, and decades of deposited structures made that breakthrough possible. Does every materials domain need the same bootstrap? Fedus says the greatest internal advances have come where data is abundant. Generalization exists among systems strongly governed by quantum mechanics, but a model that is accurate on quantum objects does not much help fluid dynamics, which sits at another level of abstraction.

On architecture, Fedus still uses language models because they are a natural interface, but he treats them as an orchestration layer: a copilot that can also direct experiments and call other specialized models. Periodic builds neural nets designed for atomic systems, with symmetry awareness, much lower latency, and fine-tuning for that domain. The orchestrator can ingest literature, experimental data, and other modalities, and it can use those specialized nets as tools and as reward functions.

Guo says she has seen the same pattern in customer support and other verticals: a general model on top, specialists underneath. Transformers remain powerful; the split of labor is what is becoming standard. Gil then presses the commercial question. Language plugged into almost all of human life. Materials science, like robotics, still has a smaller footprint. Who buys this first?

Periodic treated itself as customer zero, working closely with scientists to change how materials research is done. Fedus sees a broader market in every enterprise bottlenecked by materials or process engineering. The interfaces are familiar: engineers asking questions about their data, hunting aberrations, debugging machinery, and hunting a better formulation. Internally they built a testing ground. They are now excited enough about the technology to take the same acceleration to advanced manufacturing more broadly.

Gil asks whether this is a biotech-style model: partner with incumbents for a royalty, or keep high-value discoveries in-house. Fedus says they think of themselves as an intelligence layer, a system of record and a control plane for experiments. Some breakthroughs could be valuable enough to look like a discovery business. They are starting, though, as a software company.

Guo describes Neal Stephenson's 1990s novel *The Diamond Age*: an AI primer cloned across China, and matter pipes into homes that print whatever you download, plus evolving nanobots. What does the world look like in ten years if Periodic works? Fedus says you go from systems that write essays and software to systems that generate matter, with implications for semiconductors, aerospace, and energy. Software engineering already looks wildly different than it did six months ago. Atoms are hard, and physics sets limits, but that does not mean there is not an order of magnitude or two of speedup in making sense of data and reaching solutions faster.

What they want to give humanity, he says, is agency for atomic rearrangement and synthesis. If the physical world could keep up at some fraction of the digital world's pace, life would feel dramatically different. Guo analogizes it to a materials version of the agricultural revolution: a sudden spike in the productivity of output. The bottlenecks that have constrained that spike are the ones Periodic is trying to remove.

What excites Fedus most is the collaboration itself. The problem is irreducibly multidisciplinary: physicists and chemists working with some of the top AI researchers and engineers in the world. People who have spent decades in a field are seeing, firsthand, that under intelligent systems the work can look entirely different. That mix, he says, is not a nice-to-have. It is the job.

He analogizes to machine learning. Early Google Brain pushed the frontier with a few GPUs and a few people. Now the field is industrialized, with hundreds of researchers and hundreds of thousands to millions of GPUs, driven by scaling laws. That predictability is what let capital flood in. He thinks physical science and engineering will have a similar property once those scaling properties are established and much larger sets of experiments can be brought to bear. Intelligence and automation both matter. Improve automation and the bottleneck moves to intelligence, and scientists already feel they cannot make sense of the throughput.

Gil asks if this is as capital-intensive as frontier LLM labs. Fedus says they will need more capital. GPUs are extraordinarily expensive; compute often costs more than the physical infrastructure, which looks cheaper until you count lead times and the difficulty of well-calibrated physical systems. From a capital perspective, the bill is primarily compute. Guo notes the pay gap between a Stanford postdoc and a machine-learning engineer, and argues that many academic scientists are undercompensated relative to their societal value.

Hiring is split, loosely, into bits and atoms. On the bits side Periodic wants pre-training, mid-training, and infrastructure. On the atoms side it wants control engineering, systems engineering, and now product engineering that spans the two. Some of the scientists who have joined, Fedus says, are among the best in the world, and the chance to work at real scale is part of why they came.

On AGI and ASI, Fedus starts with a fallacy: treating intelligence as a scalar. These systems are oddly spiky. You can architect a model that is world-class in some math domain, then perturb the questions and watch it degrade to a bad high-school student. Adjacent fields can fail to generalize in non-intuitive ways. Recursive self-improvement, he says, is akin to neural architecture search from about ten years ago, and software engineering has the clearest path: huge data, cheap verifiable environments, unit tests that go from fail to pass on a few CPUs almost instantly, and no domain-expertise gap between AI researchers and software engineers.

Asked when machine self-improvement flips from human-directed to mostly automatic, he puts a domain caveat on the answer. A system that can write complete repositories, find bugs, and refactor code does not suddenly understand biology. There is a knowledge gap, and also a strategy gap: software engineering is highly verifiable, whereas scientific work is decision-making under uncertainty. In software, he says, self-improvement is happening "now-ish." AI research is next, with a slower outer loop. Instead of unit tests, you check scaling properties, whether a model converged, and how it generalizes, which takes GPUs and hours.

Both of those systems are being trained in a closed loop against their domain. That is Periodic's premise: build the same closed loops for actually doing science and engineering. The rest of the world will follow with some delay. Guo asks whether they need sufficiently good robots, something like Pi or Skild, before that loop hits escape velocity. Fedus says no, but it would be a huge accelerator. The goal is high-quantity, high-quality, diverse data; automation is assistance. People plus reliable autonomous equipment already produce large amounts of trustworthy data.

A dexterous humanoid that could wander into an unstructured lab and follow instructions would change how fast new labs stand up. Guo's own company Color had to customize liquid-handling robots with cameras, ML monitoring, and 3D-printed vibration damping because the volumes were tiny and the firmware was painful. Periodic currently uses simple, commoditized, off-the-shelf robotics and is not doing much innovation there. Asked what excites him outside Periodic, Fedus does not hesitate: robotics as the interface layer between AI and the physical world, a massive opportunity in a decade of labor shortages.